How do fine-tuned models and RAG systems power personalized AI agents that learn, collaborate, and transform enterprise workflows? What kind of technical challenges do we need to first examine before this becomes real?

As AI agents become more capable, their real-world performance increasingly depends on how well they can coordinate tools. This month's paper introduces a benchmark designed to rigorously test how AI agents handle multi-step tasks using the Model Context Protocol (MCP) — the emerging standard for tool integration. ​The authors present 101 carefully curated real-world queries, refined through iterative LLM rewriting and human review, that challenge models to coordinate multiple tools such as web search, file operations, mathematical reasoning, and data analysis.

Modern LLMs are defined as much by how they’re trained as by what they learn. This post unpacks the often-overlooked foundations of that process: pretraining—the stage that shapes a model’s core reasoning and knowledge. Starting with ULMFiT’s breakthrough in transfer learning and InstructGPT’s formalized multi-stage pipeline, it explores how pretraining has evolved into a dynamic ecosystem of techniques, from instruction-augmented and multi-phase approaches to continual and reinforcement-based pretraining. Amid the growing complexity and shifting definitions, one truth remains: understanding pretraining is essential to understanding how language models think, reason, and behave.

Unicorn Mafia won the recent hackathon at Raise Summit and explained to me what they built, including all the tech they used under the hood to make their AI agents work.

What if AI could actually remember like humans do? Biswaroop Bhattacharjee joins Demetrios Brinkmann to challenge how we think about memory in AI. From building Cortex—a system inspired by human cognition—to exploring whether AI should forget, this conversation questions the limits of agentic memory and how far we should go in mimicking the mind.

Terraform becomes messy at scale—too much duplication, manual setup, and no orchestration. Terragrunt fixes this by automating state management, reducing repetition, and handling dependencies. In 2025, its new Stacks feature enables reusable infrastructure patterns, making it the better choice for multi-environment setups despite a small learning curve.

Alex Salazar, CEO of Arcade, argues that chatbots are useless without the power to take action. He claims real value in AI lies in agents that can actually do things—trigger workflows, manage authorizations, and connect to tools like Gmail or Slack. Salazar calls out why most “agents” never make it to production—security nightmares, high costs, latency, and poor accuracy—and says Arcade fixes this by giving developers the tools to build real, authorized agents fast. He challenges the industry’s obsession with data, insisting that AI’s future is about software, not datasets, and that the heavy lifting by OpenAI and Anthropic has already changed the rules.

Your AI agent isn’t failing because it’s dumb—it’s failing because you refuse to test it. Chiara Caratelli cuts through the hype to show why evaluations—not bigger models or fancier prompts—decide whether agents succeed in the real world. If you’re not stress-testing, simulating, and iterating on failures, you’re not building AI—you’re shipping experiments disguised as products.